Autocorrect fails could be distant mammary as computers start to learn like us

Robots may soon be able to pick up new concepts more efficiently. Photo / Bloomberg

Most mobile phone users have suffered the indignity of an embarrassing autocorrect mistake, but smartphones may soon be smart enough to understand what we are trying to say.

Scientists claim that for the first time, they have programmed a machine to learn in the same way as humans.

Usually computers recognise speech and language because they have been fed hundreds of thousands of words and phrases that they keep in large datasets. But when an unknown word crops up they struggle to think outside the box, and often substitute it for something entirely inappropriate.

Humans, in contrast, need only see a new word once or twice in context to get an idea of what it means and how it can be used in different situations.

Now, in an important breakthrough for artificial intelligence, scientists have shown that it is possible for machines to learn new concepts in the same way.

So in future, rather than trying to autocorrect a strange word, a smartphone would recognise that it made sense in context and add it to its dictionary, learning like a child does.

It could also help computers recognise sign language and other gestures.

Joshua Tenenbaum, professor of cognitive science and computation at the Massachusetts Institute of Technology, said: "We believe we have made an important step here. For the first time, we think we have a machine system that can learn in ways that are hard to distinguish from human learning.

"People can learn new concepts fairly quickly, often from just a few examples, or just one. You show even a young child a horse, or school bus, and they get it from one example.

"Imagine if your smartphone could do this. You use a word, and your smartphone asks you what it means and is able to recognise the next time you are saying that to build its repertoire.

"Improving machines' ability to quickly acquire new concepts will have a huge impact on many different artificial intelligence-related tasks including image processing, speech recognition, facial recognition, natural language understanding, and information retrieval."

To show that it is possible to create machines which learn like humans, scientists programmed a computer to learn how alphabets around the world are written using just a few different examples of characters. The computer was then asked to create a new symbol based on what it knew about how the letters were formed.

Human volunteers were then asked to carry out a similar task and the results were compared to see if it was possible to distinguish between the computer-generated characters and those invented by humans.

Volunteers were unable to tell which had been created by people, showing that the machine had passed "The Turing Test", a benchmark invented by Alan Turing, the British computer scientist and codebreaker, to determine whether computers could be lifelike enough to fool humans.